Claude Code’s revised projects adds AI orchestration, but local developers must wait

Screenshot by David Gewirtz/ZDNET

ZDNET’s key takeaways

• Claude Code Projects could tame multi-agent chaos.
• Cloud users get persistent threads and shared memory.
• Local workflows remain outside the new system for now.

---

Claude Code has always had projects, in the sense that a given directory is declared as the root for source code. I keep each of the apps I’m developing in separate directories, and I’ve built up a complex orchestration and management workflow that allows certain messages to be exchanged between these directories. Beyond that, they’re isolated code structures.

Also: After the vibe-coding rush comes the debugging hangover

Today, Anthropic is announcing that it’s redesigning the idea of projects in Claude Code, adding both superpowers and hard limits. For certain classes of work, they can be game-changing. For other types of work, they’re essentially irrelevant.

Right now, this capability is being offered to “Select Claude Pro and Max subscribers who use cloud sessions in Claude Code and don’t have any existing projects on the web and desktop,” Anthropic said. Features will roll out to other users over time, but the company said that “If you are a Pro and Max subscriber already using projects, they stay as they are for now, and we’ll upgrade them later.”

That rules me out, because all of my Claude Code-related development is being done in terminal and on local files. But for those living in the cloud, the new projects capability brings life to those structures.

Anthropic said, “Claude scopes the request, delegates the work, coordinates parallel threads, reviews the outputs, and assembles the finished result. You can steer progress throughout, and it keeps working after you close your laptop.”

Features for cloud coders

One of the more pressing challenges when launching a bunch of agents is keeping them all on track. Imagine trying to get 20 oh-so-fluffy kittens to go through a little door in alphabetical order according to their kitty names. It’s not going to happen, but it will be a cute overload of coordination failure.

Also: Treat your AI agents like eager but misguided human interns – before you lose control

Coding agents are like that, but a lot less cute. To get agents to work together toward a goal, you need orchestration. This is basically a managerial-level coordinator that goes beyond the prompt. Claude Code’s revamped projects feature provides that orchestration.

You also get addressable threads. Each discussion becomes a full cloud session, with its own branch and copy of the repo it’s working on. The coordinator provides instructions that route sessions into existing threads as appropriate.

Anthropic advised you to brief Claude on a project “the way you’d brief a chief of staff, and it routes work to new or pre-existing threads.”

Because those threads are named, you can steer them mid-flight. You can dig into a running thread and redirect it as necessary. No more having to wait for it to run to completion, wasting your time and consuming tokens, only to produce something you can already see is going wrong.

Also: How to install and configure Claude Code, step by step

Threads can be aware of each other. Anthropic said, “Over time, Claude learns more about the project details and applies them to its work. Every thread now adds to and draws from a shared memory, reducing the need for complex prompt engineering.”

This is a form of automated memory that can prove quite powerful in the right context. It allows your work to span repos, so you can work across several repos at once, work with a thread tied to each, and then get feedback on which pull requests need to be merged in what order.

And since it’s all in the cloud, if your computer is shut down, you’ll be able to let it keep running. Anthropic added that mobile check-ins are on the way.

What’s not new

This doesn’t become some sort of new parallelism. I’ve been using swarms of subagents for a while now. They can fan out and do work. In fact, Anthropic’s announcement said that this architecture is what the new projects use internally.

Also: 75% of developers I surveyed prefer Claude Code over Codex – here’s why

Claude has also been able to split work across subagents. The difference is that the split persists, and you can go back into the discussion for each subagent and chat about it.

One other thing to note: this feature doesn’t appear to be strictly autonomous. It looks like threads begin when you hit the Start button for each thread Claude recommends.

As mentioned above, at least for now, this is not yet a tool for local developers like me. But Anthropic added that “Threads run in the cloud today; running on your machine alongside your local tools and code and behind your network is coming very soon.”

Also, Anthropic isn’t taking away the terminal. This is an additional feature. Nothing the company described in its blog post indicates any negative changes to local or CLI functionality.

Putting it all in perspective

While I was writing this, I stopped and ate a burger. While munching, I thought about what this new capability means and came up with three ideas worth considering. Here’s a lightning round of things to consider when evaluating this capability.

• It remembers, but what does that mean? Memory isn’t well-documented in Anthropic’s post. There is no apparent way to inspect what’s been kept, correct it when it goes wrong, export it, nor are there any details about how long it’s retained.
• Cost and usage scales as you fan out more agents. This is the case now, with subagents, but it’s worth reinforcing. Anthropic said, “Projects can run several threads at once, and each one is a full Claude Code session. Because of this, projects reach usage limits faster.” I’ve definitely been cut off after running a fleet of agents across my code.
• Conflicts are pushed down the road; they’re not prevented. Threads that overlap will surface as merge conflicts that you’re going to need to resolve by hand.

Questions for Anthropic

I have many, many, oh-so-many questions for Anthropic about this, especially once some of these capabilities become available to desktop users. I’ve narrowed them down to three worth thinking about, at least for now.

1. If a project outgrows its context (which mine do every few hours or so), what happens to that shared memory? Can a user see what was kept? Do they all wind up as .md files somewhere?
2. Is a project’s history automatically maintained, and can it be exported with prompts intact? My memory management system requires Claude to record every prompt I give it, which provides both an engineering audit path and something of a scientific journal for evidence of conception, diligence, reduction to practice, and inventorship.
3. When the capability to use this feature on local projects becomes available, will projects inherit memory, permissions and hook behavior? How will it work with existing memory management systems and existing, deeply evolved projects?

From Anthropic’s specific statement that this feature is for those who “don’t have any existing projects on the web and desktop,” I’m guessing migration is not yet resolved.

If you could steer a running Claude Code thread before it goes off the rails, how much would that change the way you use coding agents? Let us know in the comments below.

---

You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.